Unified Video Dense Prediction from Disjoint Data

Researchers introduced UniD, a unified video model for eight dense prediction tasks, including depth, normals, segmentation, boundaries, human parts, albedo, shading, and materials.

UniD learns from fragmented, task-specific datasets by distilling knowledge from lightweight task experts instead of requiring fully annotated multi-task data.

The model uses pretrained diffusion priors as a backbone and shows strong generalization across tasks and video scenes.

This approach reduces reliance on expensive pseudo-labeling and demonstrates a practical path to unified video understanding from disjoint data.